Midjourney v6: Photorealism Specs, Pricing & Discord Workflow (2026)

Midjourney v6 is the most influential image generation model you can’t actually deploy. Released December 21, 2023, it redefined photorealistic AI art through a Discord bot interface, achieving an FID score of 8.2 and winning 73% of human preference tests against DALL-E 3 for realistic portraits. But there’s no API, no model weights, no technical documentation. Just a $10-to-$120 monthly subscription and a chat command that spits out better product photography than most professionals can shoot.

This creates a paradox for 2026. Midjourney v6 is technically obsolete (v7 launched February 2026 with faster inference and native video). Flux.1 Pro beats it on benchmarks. Stable Diffusion 3 costs less. Yet 60% of Midjourney’s paying users still force v6 via the –v 6 flag because nothing else nails photorealism the same way.

The reason matters if you’re evaluating AI image generation for commercial work, creative projects, or competitive analysis. Midjourney v6 isn’t a model in the traditional sense. It’s a service built on a closed diffusion architecture that prioritizes output quality over everything else: developer access, enterprise integration, transparency, cost efficiency. The iPhone of AI art, as one developer put it on Hacker News. Expensive, locked down, and still the thing everyone compares their work against.

This guide covers how v6 actually works, what you can do with it, and why its closed architecture makes it fundamentally different from every other AI tool covered on this site. If you’ve been generating images with DALL-E or Stable Diffusion and wondering whether Midjourney’s hype is real, the answer is yes and also it’s complicated.

Midjourney v6 runs on proprietary infrastructure you’ll never touch

The core frustration with Midjourney v6 is that nobody outside the company knows exactly what it is. No model card. No architecture paper. No parameter count. The best guess from reverse-engineering Discord API responses and inference timing suggests an 8-to-12 billion parameter diffusion transformer, roughly SDXL scale, but that’s speculation.

What Midjourney has confirmed: v6 uses a custom DiT (Diffusion Transformer) architecture evolved from Stable Diffusion’s U-Net backbone, with multi-head attention layers trained specifically for photorealistic textures and lighting. The training data is billions of image-text pairs scraped from the internet, LAION-style, with manual curation favoring art and photography over diagrams or technical illustrations. Data cutoff is approximately mid-2022.

Access happens two ways. The Discord bot (type /imagine followed by your prompt in any Midjourney server channel) or the web interface at alpha.midjourney.com. Both require a paid subscription. Both hit the same cloud infrastructure, likely AWS US-East based on latency tests. There is no third option. No local deployment, no API endpoint, no model weights to download.

The subscription model bills per GPU-hour, not per image. Standard tier gives you 15 hours of “fast mode” generation per month, which translates to roughly 1,800 images if you’re generating at default settings. After that, you drop into “relax mode,” where generation is unlimited but queues stretch to 10-15 minutes per image during peak hours. Pro and Mega tiers add more fast hours and priority queue access.

This pricing structure makes cost-per-image calculations weird. At $30 per month for Standard, a typical 1024×1024 generation costs about $0.017 if you max out your fast hours. That’s cheaper than DALL-E 3’s $0.04 per image. But if you need 10,000 images for a commercial project, you’re stuck waiting in relax mode or upgrading to Mega tier at $120 monthly, which still caps fast generation at 60 hours (roughly 7,200 images). There’s no bulk discount, no enterprise API with per-image pricing. Just subscriptions.

The 30-second inference time is 2× faster than v5 but slower than Flux.1 Pro (25 seconds) or local SDXL on an RTX 4090 (15 seconds). The quality improvement justifies the wait for most users, but if you’re iterating on 50 variations of a logo design, those 30-second delays add up to 25 minutes of dead time.

V6 beats DALL-E 3 on photorealism, loses to Flux.1 on everything else

Midjourney v6 launched with one goal: make AI-generated images indistinguishable from professional photography. The FID score of 8.2 (lower is better) represents a 32% improvement over DALL-E 3’s 12.1 and a 47% improvement over Stable Diffusion XL’s 15.4. That’s measured improvement in how closely generated images match the statistical distribution of real photos.

But Flux.1 Pro, released mid-2024, hit 7.9. And v7, Midjourney’s own successor, clocks in at 9.1, slightly worse than v6. This explains why users keep forcing v6 via the –v 6 parameter despite v7 being the new default.

The CLIP score of 0.342 measures how well v6 follows text prompts. It beats DALL-E 3 by 4% and SDXL by 8%, which sounds small until you’re generating images with complex spatial relationships like “red car parked in front of blue house with green door.” V6 gets that right 73% of the time in testing. DALL-E 3 manages 61%. SDXL struggles at 48%.

Where v6 dominates: photorealistic portraits, product photography, architectural renderings, anything requiring accurate human anatomy and realistic lighting. The 87% hand and face accuracy beats every 2023-era model. In blind tests on Midjourney’s Discord arena, users identified v6 outputs as “real photos” 41% of the time versus 23% for DALL-E 3.

Where v6 falls short: abstract art, illustrations, graphic design, anything non-photorealistic. The model over-applies photorealistic textures even when you ask for flat colors or cartoon styles. Flux.1 and DALL-E 3 handle illustration prompts better. V6 also can’t render text in images, producing gibberish 98% of the time. If you need a sign, logo, or label in your image, you’re adding it in Photoshop afterward.

Speed matters for iteration. At 30 seconds per image, generating 20 variations to find the right composition takes 10 minutes. DALL-E 3 at 45 seconds stretches that to 15 minutes. Local SDXL at 15 seconds cuts it to 5 minutes. But the quality gap means most professional users tolerate v6’s speed because the first result is more likely to be usable.

The cost comparison gets messy. V6 at $0.04 per image (Standard tier average) matches DALL-E 3’s API pricing. But DALL-E 3 offers batch discounts and enterprise contracts. V6 offers neither. Local SDXL costs $0.02 per image after hardware investment (RTX 4090 at $1,600), which breaks even after 80,000 images. For most users, that’s never. For production studios generating 500+ images daily, local deployment wins on cost.

Multi-stage latent upsampling makes photorealism actually work

The technical innovation that separates v6 from competitors is how it generates high-resolution images without the blurriness or artifact patterns that plague other diffusion models. Midjourney calls it “multi-stage latent upsampling with attention refinement.” In practice, it’s a three-pass generation process: rough sketch at 512×512, detail layer via cross-attention, final upscale to 1024×1024 or 2048×2048 with a dedicated super-resolution module.

Most diffusion models, including SDXL and DALL-E 3, generate images in a single pass at target resolution or use post-processing upscalers like RealESRGAN. That works for composition but struggles with fine texture detail. Skin pores, fabric weave, water reflections, the stuff that makes photorealism convincing, those require training data at high resolution and architecture that can synthesize detail without hallucinating nonsense patterns.

V6’s approach trains separate diffusion stages on different resolution datasets. The coarse generation learns composition and color from lower-resolution images (faster training, broader dataset). The detail inpainting stage trains specifically on high-resolution photography, learning texture patterns that correspond to real-world materials. The upsampling module uses a dedicated super-resolution network, not a generic upscaler.

Proof: in blind A/B tests with 1,200 participants on Midjourney’s Discord in January 2024, v6 images generated with the –q 2 parameter (which activates additional refinement passes) were identified as “real photos” 41% of the time. DALL-E 3 managed 23%. SDXL hit 18%. The FID score improvement of 28% over v5 correlates directly with the new upsampling architecture, according to v6’s release notes.

The –q 2 parameter trades 40% more compute time (45 seconds versus 30 seconds) for measurably better texture coherence. Use it for final outputs, client presentations, anything where quality matters more than iteration speed. Stick with –q 1 (the default) when you’re exploring compositions or testing prompt variations.

This feature doesn’t work miracles. It won’t fix a bad composition or make text legible. But for photorealistic subjects, natural environments, product shots, anything where surface detail sells the realism, v6’s upsampling architecture is the reason it still holds the photorealism crown 27 months after launch.

Commercial work drives 70% of Midjourney’s actual usage

Product photography without the photoshoot

E-commerce brands use Midjourney v6 to generate lifestyle product shots that would cost $2,000-$5,000 per day of studio time. A furniture company creates 50 room mockups in 2 hours: living rooms with different wall colors, lighting conditions, decor styles. Total cost: $6 for 150 images on a Standard subscription. The 87% hand and face accuracy matters here because lifestyle shots often include people interacting with products. The CLIP score of 0.342 ensures the sofa actually matches the product description.

This workflow killed an entire tier of commercial photography. Photographers report 40% revenue decline since Midjourney v6 launch, according to a 2024 industry survey, as brands shift budgets to AI-generated imagery that costs 95% less than traditional shoots.

Concept art for film and gaming pre-production

Pre-production teams generate 100+ environment concepts per day for client review. Used on productions including Netflix originals (unconfirmed but widely reported by Discord users working in VFX). The human preference Elo of 1250 for “cinematic environments” means v6 outputs look professional enough for pitch decks and mood boards. The –ar 16:9 parameter produces widescreen compositions that match film aspect ratios.

VFX artists now compete with Midjourney-trained generalists who can produce concept art at 10× the speed, fundamentally changing studio hiring practices in 2024-2026.

Social media content at scale

Influencers and brands generate Instagram and TikTok visuals without hiring photographers. Typical workflow: 20 prompts, 80 variations, 5 final posts. The 73% user preference for v6 photorealism on “realistic person” prompts makes it the default for lifestyle content creators. The –sref parameter enables consistent brand aesthetics across posts, maintaining visual identity across hundreds of images.

While video tools like Seedance 2.0 dominate short-form content, Midjourney v6 remains the standard for static social media imagery, with an estimated 2 million images generated daily for Instagram alone.

Architectural visualization before construction

Architects create photorealistic renderings of unbuilt structures for client presentations, replacing traditional 3D rendering workflows that take 4-8 hours per image. The LPIPS score of 0.21 indicates high perceptual similarity to reference images. The –iw parameter allows image-to-image refinement from sketches, letting architects upload hand-drawn concepts and get photorealistic versions in 30 seconds.

Real estate developers now use Midjourney v6 for property marketing before construction begins, but legal disputes over “misleading” AI renderings have increased 340% since 2024.

Book covers and editorial illustration

Publishers generate cover concepts in minutes versus weeks of back-and-forth with illustrators. Self-published authors create professional covers for $10 monthly subscription instead of $500-$2,000 per traditional cover. The 94% style consistency with –sref parameter means series books can maintain visual coherence across 10+ volumes.

The explosion of AI-generated book covers correlates with a 600% increase in self-published titles since Midjourney v6 launch, overwhelming library acquisition systems.

Character design with consistent appearances

Game developers and comic artists maintain consistent character appearances across hundreds of images using the –cref (character reference) parameter. Testing shows 89% facial consistency across 50-image sequences, critical for visual novels and webcomics where the same character appears in different scenes, poses, and lighting conditions.

Virtual influencers like Lolita Cercel rely on Midjourney v6’s character consistency features to maintain recognizable appearances across thousands of social media posts.

Medical and scientific visualization

Researchers create anatomical illustrations, molecular structures, and educational diagrams without hiring medical illustrators. The photorealism score of 9.1 out of 10 makes outputs suitable for textbooks and online courses. But accuracy concerns limit use to educational contexts, not clinical decision-making, because the model can hallucinate anatomical details that look realistic but are medically incorrect.

While Claude for Healthcare targets clinical workflows, Midjourney v6 dominates medical education content creation, generating an estimated 50,000+ anatomical illustrations monthly for textbooks and online courses.

Fashion and apparel mockups

Clothing brands visualize designs on models without physical samples or photoshoots, enabling rapid A/B testing of colorways and styles. The 87% hand and face accuracy is critical because fashion requires realistic body proportions and natural poses. The –ar 2:3 parameter produces portrait-oriented fashion shots that match industry standards.

Fashion content creators now generate entire lookbooks using Midjourney v6, with AI-generated outfit inspiration posts receiving 3× more engagement than traditional photography on Pinterest.

Discord commands are the only interface that exists

Midjourney v6 has no official API. Every integration you see in the wild uses Discord API wrappers to automate bot interactions, which violates Midjourney’s Terms of Service if used for commercial automation. The company has never released REST endpoints, Python SDKs, or OpenAI-compatible interfaces.

The primary method is typing /imagine in any Midjourney Discord channel, followed by your prompt and parameters. The bot processes your request, queues it, and posts the result as a Discord message with four image variations. You click buttons to upscale individual variations or generate new sets.

The web interface at alpha.midjourney.com offers the same functionality through a cleaner UI. You type your prompt, select v6 from a version dropdown, adjust parameters via sliders, and download results directly. But under the hood, it’s hitting the same infrastructure as the Discord bot.

Parameters control every aspect of generation. The –v 6 flag forces version 6 (required as of February 2026, since v7 became the default). The –ar parameter sets aspect ratio: –ar 2:3 for portraits, –ar 16:9 for landscapes, –ar 1:1 for squares. The –q parameter controls quality: –q 1 for fast iteration, –q 2 for final outputs with 40% longer generation time but better texture detail.

Stylization ranges from 0 to 1,000 via the –s parameter. Lower values produce literal interpretations of your prompt. Higher values add artistic flourishes. The sweet spot for photorealism is –s 100-250. Above 750, you get unwanted painterly effects. The –c parameter controls chaos (variation), ranging 0-100. Set it to 0 for consistent results across re-rolls, 50-80 for diverse options, 100 for experimental outputs.

Image prompts work via the –iw parameter. Upload a reference image to Discord, copy its URL, paste it before your text prompt, and add –iw 1.5 to weight the image influence. Higher values (up to 2.0) make output closer to the reference. Style references use –sref with a style image URL and –sw to control style strength (0-1000).

Character references maintain consistent faces via –cref with a reference image and –cw for character weight (0-100). A weight of 100 means exact facial match, 0 means loose inspiration. This is critical for maintaining character consistency across series.

Negative prompts don’t exist in Midjourney’s syntax. Instead, use the –no parameter: –no hands, text, watermark. It’s less precise than Stable Diffusion’s negative prompt system but works for common exclusions.

For developers trying to automate workflows, the unofficial workaround involves discord.py to monitor bot messages and extract image URLs. This is unreliable, rate-limited, and against TOS. The recommended approach for commercial projects is manual generation through the web interface, which at least provides a cleaner download workflow than screenshotting Discord messages.

Check the official documentation for parameter syntax and examples, since Discord bot commands are the only interface that will ever exist.

Prompt engineering for v6 requires unlearning Stable Diffusion habits

Midjourney v6 truncates prompts after roughly 80 words. Longer prompts don’t get summarized or compressed, they just ignore everything past the cutoff. Test this by generating the same image with a 120-word prompt and an 80-word version. The outputs will be identical because the model never saw the final 40 words.

Comma-separated structure works best. “Subject, environment, lighting, style, technical details” yields more consistent results than paragraph-form descriptions. Example: “software engineer at desk, modern office, natural window lighting, shallow depth of field, shot on Canon EOS R5, 85mm f/1.4, photorealistic” hits all the elements v6 prioritizes without wasting tokens on connecting words.

Weight syntax doesn’t exist. Stable Diffusion users expect (word:1.5) to emphasize terms. V6 ignores that syntax entirely. Instead, repeat important words or use natural language emphasis: “extremely detailed face, highly detailed skin texture” works better than trying to hack weights.

The –q parameter makes a bigger difference than most users realize. Quality 1 (default) generates in 30 seconds with good results for iteration. Quality 2 adds 15 seconds but improves texture coherence by roughly 15% based on FID scores. Use –q 1 when exploring compositions, –q 2 for anything client-facing or final.

Stylization between 100-500 covers most use cases. Below 100, v6 produces overly literal results that look flat. Above 750, it adds artistic flourishes you didn’t ask for (painterly brushstrokes on photorealistic portraits, dramatic color grading on product shots). The default of 100 works for most prompts. Push to 200-300 for more dramatic lighting or composition.

Chaos controls variation across the four-image grid. At –c 0, all four variations look nearly identical (useful when you’ve nailed the prompt and want minor differences). At –c 50-80, you get diverse interpretations of the same concept (useful for exploring options). At –c 100, outputs can vary wildly in composition and style (useful for experimental work but frustrating when you want consistency).

Aspect ratio affects composition logic, not just crop. The –ar 16:9 parameter triggers landscape framing with horizon lines and wide establishing shots. The –ar 2:3 parameter triggers portrait framing with vertical emphasis and tighter crops. Square (1:1) often produces awkward compositions because v6’s training data skews toward standard photo aspect ratios.

For photorealism, include camera specs in your prompt. “Shot on Canon EOS R5, 85mm f/1.4, natural lighting, shallow depth of field” activates training patterns associated with professional photography. This isn’t magic, it’s pattern matching against billions of images tagged with EXIF data.

For consistent characters across images, use –cref with a reference image URL and –cw 100 for exact facial match. Generate your reference image first with a detailed character description, upscale it, then use that URL in subsequent prompts. Testing shows 89% facial consistency across 50-image sequences when using –cw 100.

For style consistency across series, use –sref with a style reference image and –sw 500 for balanced style transfer. Lower weights (100-300) give subtle style hints. Higher weights (700-1000) force aggressive style matching that can override your text prompt.

Common failure modes: text in images produces gibberish 98% of the time (add text in Photoshop afterward). Complex spatial relationships like “Person A standing behind Person B holding object C” fail 40% of the time (use –iw with a reference sketch instead). Specific counts like “exactly 5 apples” produce 3-7 apples randomly (generate 10 variations, pick the one that matches).

Banned content gets filtered inconsistently. Gore, violence, and nudity are blocked by default (stricter than DALL-E 3). Celebrity names sometimes work, sometimes get blocked (no public documentation of the rules). Copyrighted characters (Disney, Marvel) sometimes generate, sometimes fail (enforcement is inconsistent). Medical imagery (surgery, anatomy) works for educational use but may trigger filters if the prompt sounds clinical.

Text rendering fails, Discord workflow blocks enterprise adoption, queue times punish free users

Midjourney v6 cannot generate legible text in images. Attempts to include signs, labels, or typography produce random letter-like shapes 98% of the time. This is a known limitation acknowledged in official documentation. Competitors like DALL-E 3 and Ideogram achieve 80%+ text accuracy. There’s no workaround except generating the image, then adding text in Photoshop.

The Discord-only workflow blocks enterprise adoption. No API means IT departments can’t track usage across teams, enforce data governance policies, integrate into existing creative pipelines, or operate in air-gapped environments. A survey of 200 enterprise creative teams in May 2024 found 73% cited “lack of API” as the primary barrier to adoption.

Content filtering is inconsistent. The moderation system blocks some celebrity names but allows others, filters medical imagery inconsistently, sometimes flags abstract art as “gore.” There’s no public documentation of filtering rules. Midjourney staff responses in Discord confirm “evolving moderation system” but provide no specifics.

Queue wait times on free and Basic tiers make relax mode nearly unusable during peak hours. Relax mode (unlimited slow generation) takes 10-15 minutes per image when US and EU users are both active. Fast mode quota on Basic tier exhausts quickly: 200 images per month equals roughly 6 images per day. Power users hit the limit in the first week.

Training data bias skews toward Western aesthetics. Prompts for non-Western subjects often produce stereotypical or inaccurate results. An academic analysis by the Stable Diffusion Bias Research Group in August 2024 found 68% Western-centric outputs regardless of prompt specificity when testing 1,000 diverse prompts.

Batch processing doesn’t exist. You can’t generate 100+ images from a CSV of prompts. Each image requires manual Discord command or web interface interaction. Competitors like Stable Diffusion and DALL-E 3 offer batch APIs that process thousands of images overnight.

Prompt drift on complex scenes means multi-object compositions ignore 30-40% of prompt details. Example: “red car in front of blue house with green door” might produce blue car, red house, no door. The CLIP score of 0.342 is good but not perfect. User testing shows 35% of complex prompts require 3+ re-rolls to match intent.

Copyright and legal ambiguity persists. Training data sources are undisclosed. Potential copyright issues with generated outputs. Several lawsuits filed against Midjourney by artists (2023-2024, some settled, terms undisclosed). Legal filings in Andersen v. Stability AI name Midjourney as a defendant. The company’s Terms of Service grant users rights to outputs but don’t clarify training data provenance.

No certifications, public-by-default Discord images, ongoing litigation

Midjourney stores all prompts and generated images on company servers for 30 days (Standard and Pro tiers) or indefinitely (Basic tier unless manually deleted). Users can disable use of their generations for model training via account settings (added March 2024), but there’s no way to prevent storage entirely.

All Discord-generated images appear in public channels unless you pay $20 extra monthly for “Stealth Mode” (Pro and Mega tiers only). Images generated via the web interface are private by default but still stored on Midjourney servers. There’s no option for client-side generation or encrypted storage.

The company has published no SOC2, ISO 27001, or HIPAA certifications as of March 2026. No third-party security audits appear in public documentation. The privacy policy mentions EU users but doesn’t specify GDPR-compliant data handling procedures. COPPA compliance is questionable since Discord requires users 13+ but Midjourney adds no additional age verification.

Servers run on AWS US-East (Northern Virginia) based on Discord latency tests and user reports. There are no regional data residency options. All data processes in the US regardless of user location. Midjourney is accessible in China via VPN (Discord is blocked, but the web interface works), with no official China deployment.

Enterprise tier exists with custom pricing for teams of 100+ users. It includes priority queue and dedicated support but remains cloud-only. No private cloud deployment, no on-premises installation, no VPC options. Terms of Service explicitly disclaim uptime guarantees. Service outages in 2024-2025 ranged from 30 minutes to 4 hours.

Ongoing litigation includes a class-action lawsuit by artists alleging copyright infringement in training data (Andersen v. Stability AI et al., filed January 2023, ongoing as of March 2026). Midjourney’s Terms of Service grant users full commercial rights to generated images but don’t address training data provenance. CEO David Holz stated “we used a lot of images from the internet” in an August 2023 interview with The Verge.

No major security breaches have been reported as of March 2026. But the service’s dependency on Discord infrastructure means Discord outages directly impact Midjourney availability. Five incidents occurred in 2024-2025, with the longest lasting 4 hours.

V6 launched December 2023, v7 tried to replace it and failed

Latest news

More on UCStrategies

The shift to AI-generated content extends beyond static images. OpenAI’s live video streaming API brings real-time vision to ChatGPT, enabling applications Midjourney can’t touch. For teams evaluating AI tools across modalities, understanding where each excels matters more than chasing the single “best” model.

Cost calculations get complicated when comparing subscription services to API-based tools. Google Gemini 2.0 Flash with thinking mode demonstrates how different pricing models (per-token versus per-hour) create different value propositions for different workloads.

The legal landscape around AI-generated content remains unsettled. OpenAI’s reasoning models raise similar questions about training data and copyright that Midjourney faces, but with more transparency about model behavior and limitations.

Common questions

Is Midjourney v6 still worth using in 2026, or should I switch to v7?

V6 remains superior for photorealism despite v7’s speed improvements. Discord polls show 60% of users prefer v6, and the FID score (8.2 versus 9.1) confirms better image quality. Use v7 if you need video generation or faster iteration. Force v6 via –v 6 flag for final photorealistic outputs. The quality difference is measurable and visible.

Can I use Midjourney for commercial projects?

Yes, if you have a paid subscription. Midjourney’s Terms of Service grant full commercial rights to generated images on all paid tiers (Basic, Standard, Pro, Mega). Free trial images cannot be used commercially. But legal ambiguity around training data means some clients may require indemnification clauses in contracts.

How does Midjourney v6 compare to DALL-E 3?

V6 beats DALL-E 3 on photorealism (FID score 8.2 versus 12.1) and costs the same per image ($0.04). DALL-E 3 offers an API, batch processing, and better text rendering. Choose v6 for product photography and portraits. Choose DALL-E 3 for integration into automated workflows or when you need text in images.

Can I run Midjourney locally?

No. Midjourney has never released model weights or local deployment tools. The service is exclusively cloud-based via Discord or web interface. For local deployment, use Stable Diffusion XL or Flux.1 instead. SDXL-based “Midjourney-style” fine-tunes exist on Civitai but are illegal per Midjourney’s Terms of Service.

Why can’t Midjourney generate text in images?

Text rendering requires architecture specifically trained on text-image relationships. Midjourney v6’s diffusion model treats text as visual patterns, not semantic content, resulting in gibberish 98% of the time. DALL-E 3 and Ideogram solve this with dedicated text encoding modules. The workaround is generating the image, then adding text in Photoshop or Canva.

Is my data safe with Midjourney?

Midjourney stores all prompts and images on company servers. Discord-generated images are public by default unless you pay for Stealth Mode ($20 extra monthly on Pro/Mega tiers). The company has published no SOC2, GDPR, or HIPAA certifications. For sensitive projects requiring data governance, this is a dealbreaker. For general creative work, the risk is low but the transparency is poor.

What’s the cheapest way to use Midjourney?

Basic tier at $10 per month gives 200 fast-mode images plus unlimited relax-mode generation. That’s $0.05 per fast image or effectively free if you tolerate 10-15 minute queue times. For professional use, Standard tier at $30 per month provides 15 GPU hours (roughly 1,800 images), working out to $0.017 per image. Heavy users need Pro ($60) or Mega ($120) to avoid constant queue waits.

Can Midjourney generate consistent characters across multiple images?

Yes, using the –cref parameter with a reference image and –cw 100 for exact facial match. Testing shows 89% facial consistency across 50-image sequences. Generate your reference image first with detailed character description, upscale it, then use that URL in subsequent prompts. This works for faces and overall character design but struggles with specific clothing or accessories.